Efficient opinion sharing in large decentralised teams

Efficient Opinion Sharingin Large Decentralised Teams

Oleksandr Pryymak, Alex Rogers and Nicholas R. Jennings{op08r,acr,nrj}@ecs.soton.ac.uk

University of SouthamptonAgents, Interaction andComplexity Group

6 June 2012AAMAS'12

Disaster response and Large Decentralised Teams

2010, Haiti earthquake

Citizen and public news reporting (Ushahidi)

2010, Chile earthquake

"Twitter is one of the speediest, albeit not the most accurate, sources of real-time information" France24

Disaster response and Large Decentralised Teams

Teams are large Decentralised Few opinion sources Observations are uncertain and conflicting Agents share only opinions without supporting information (Communication is strictly limited)

Opinion is a subjective belief about the common subject of

interest

Challenge

How to improve the accuracy of shared

opinions?

Opinion Sharing Model Networked team

Opinions are introduced gradually

Noisy

Weights (levels of importance) define sharing process

Agents' model

Agents' model

Agents' model

Dynamics of the Opinion Sharing

Stable Transition Unstable

Stable Dynamics

Unstable Dynamics

Transition

Dynamics of the Opinion Sharing

Problem How to find the settings for improved reliability?

Requirements: Decentralised On-line Adaptive (i.e. complex topology, size, degree) Minimise communication

DACOR algorithm Distributed Adaptive Communication for Overall Reliability by R. Glinton, P. Scerri, and K. Sycara

introduces excessive communication overhead (#neighbours2)

exhibits low adaptivity (3 parameters to tune)

Autonomous Adaptive Tuning (AAT)

Finds tcritical

for each agent individually

Each agent must use the minimal importance level

that still enables it to form its opinion

AAT: sample run

AAT: stages

Executes 3 stages by each agent:

Select candidate importance levels

Estimate the awareness rates they deliver

Select the best one to use

However, the agent's choice highly influences others

AAT: Candidate Importance Levels

This stage limits the search space.

Initialise an agent once with candidates:

drawn from the range with a given step size. However,

the algorithm becomes computationally expensive

that lead to opinion formation on different update steps. Thus, the agent exhibits different dynamics.

AAT:Estimation of the Awareness Rates

Awareness Rate is a probability of forming an opinion with a given importance level.

2 evidences indicate that agent could have formed an opinion with a given candidate:

If an opinion was formed, then all higher levels would have led to opinion formation

Otherwise, a candidate requires less updates to form an opinion than was observed

AAT:Strategy to Choose an Importance Level

Since an agent's choice influences others, strategies with less dramatic changes to the dynamics perform better

Hill-climbing: Select the importance level which is closest to the currently used

(with the awareness rate closest to the target)

Outperforms popular MAB strategies.

Results: Target Awareness Rate

Compromise awareness for overall Reliability

Results: Target Awareness Rate

Compromise awareness for overall Reliability

Results: Reliability and Convergence

Random Network

Results: Reliability and Convergence

Scale-free Network

Results: Reliability and Convergence

Small-world Network

Results: Communication Expenses

Minimal Communication = #messages to share an opinion in a single cascade (total #neighbours)

Results: Indifferent AgentsWhat if some of the agents cannot alter their weights?

Summary

Presented a novel algorithm, AAT, that:

improves the reliability of the opinions outperforms the existing algorithm, DACOR, and prediction of

the best setting (Av.Pre-tuned) the first that minimises communication to opinion sharing only Computationally inexpensive Adaptive, scalable and robust to the presence of indifferent

agents Operates without a knowledge of the context and the ground truth

What to take away?